WaveStyleGAN
Overall design (FFHQ config as a reference):
- Generator with skip connections
- Discriminator with skip connections
- Wavelets in generator and discriminator (trainable, adjusted scales and learning rate)
- Reduced complexity of a Mapping network (3 FC layers, smaller dimension of a latent input)
- SiLU activation in generator (Synthesis and Mapping networks)
- Usage of an updated Fast Fourier Convolution blocks (FFCv2) with wavelets for Discriminator
- Differential data augmentations (DIffAug)
- ...
Generator architecture:
Discriminator architecture:
Updated FFCv2 block:
A model trained on FFHQ dataset for 9.99M images (configs/FFHQ_v1_best.yaml). Full resolution samples can be found in generated_sample directory
(samples here are downsampled to speed-up page loading)
Random samples:
Best samples:
Some curated samples with truncation (level=0.75):
